JP2014123011A - Noise detector, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect various kinds of sudden noise without increasing processing loads of equipment.SOLUTION: A noise detector comprises: a feature variation calculation part which calculates feature variation which is variation of feature quantities between two temporally adjacent frames based on either feature quantities of amplitude feature quantities or frequency feature quantities held at a holding part which holds the amplitude feature quantities and the frequency feature quantities for a plurality of frames; a section specification part which specifies a section of frames in which weighting averaging of the amplitude feature quantities and the frequency feature quantities held at the holding part should be performed by comparing the feature variation with a preset threshold; a feature quantity set generation part which generates a set of weighting average values of each of the amplitude feature quantities and the frequency feature quantities corresponding to each of the frames of the specified section as a feature quantity set; and a noise determination part which determines whether or not the latest frame of an input signal is a frame including non-constant noise which is sudden noise based on the feature quantity set.

Description

  The present technology relates to a noise detection apparatus and method, and a program, and in particular, a noise detection apparatus and method that can detect various sudden noises without increasing the processing load on the device, and Regarding the program.

  Recording devices such as IC recorders, smartphones, and video cameras record surrounding sounds using a small built-in microphone.

  When recording with such a recording device, the operation sound when the user operates the recording device using the operation buttons or the operation sound of the keyboard generated at a position separated from the recording device is recorded as noise. It will be mixed.

  Therefore, a technique has been proposed for detecting and reducing special noise such as keyboard operation sounds generated at remote positions mixed as noise during recording in a recording device (see, for example, Patent Document 1). .

  In the noise detection method of Patent Document 1, a keyboard operation sound generated at a position separated from a recording device is mainly detected.

  The operation sound of the keyboard generally appears as a set of pulsed noise signals having a relatively long duration on the recorded voice signal. For this reason, the amplitude value (signal level) of a pulse-like noise signal having a relatively long duration is compared with a threshold value, or a high frequency region component that is hardly present in an audio signal is compared with a threshold value. Noise can be easily detected.

  In addition, a technique for determining whether an input signal is voice (for example, conversation) or non-voice has been proposed (see, for example, Patent Document 2). For example, it is possible to recognize a frame determined as non-speech using the technique of Patent Document 2 as noise.

JP 2012-027186 A JP 2009-251134 A

  However, the noise recorded by the recording device is not necessarily a signal whose frequency characteristic resembles a pulse signal, such as the operation sound of a keyboard. There is a lot of noise. Such noise has been difficult to detect by conventional techniques such as Patent Document 1, for example.

  Further, many of the sudden noises recorded by the recording device (for example, long-time applause, cough, sneeze) are unstable in duration and become a value with a large variance that is almost unpredictable. It is also difficult to detect by the detection method using the attenuation feature amount, which is one of the noise detection methods according to the above technique.

  Furthermore, in the detection method using the attenuation feature amount as in the technique of Patent Document 1, since the signal is analyzed in a relatively long time range, there is a problem that a delay corresponding to the time range occurs.

  Moreover, the technique of patent document 2 is a method of determining whether an input signal is a voice to the last, and does not aim at the detection of noise. For example, even if noise detection is performed using the technique of Patent Document 2, it cannot be determined whether or not the noise is sudden noise.

  Further, the method described in Patent Document 2 is complicated in calculation, and is considered difficult to implement in, for example, a mobile device.

  The present technology is disclosed in view of such a situation, and allows various sudden noises to be detected without increasing the processing load of the device.

  One aspect of the present technology is an amplitude feature amount calculation unit that calculates an amplitude feature amount in a waveform of a predetermined frame of an audio input signal; a frequency feature amount calculation unit that calculates a frequency feature amount in the waveform of the predetermined frame; , The amplitude feature quantity and the frequency feature quantity are temporally adjacent based on any one of the amplitude feature quantity and the frequency feature quantity held in a holding unit that holds a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two frames, and a feature change amount that is held in the holding unit by comparing the feature change amount with a preset threshold value. A section identifying unit that identifies a section of a frame that is a weighted average of the amplitude feature quantity and the frequency feature quantity, and that is identified in time. A feature value set generation unit that generates a set of weighted average values of the amplitude feature value and the frequency feature value corresponding to each of the frames in the section as a feature value set; and the input based on the feature value set The noise detection apparatus includes a noise determination unit that determines whether or not a latest frame of a signal is a frame including non-stationary noise that is sudden noise.

  The amplitude feature amount calculation unit or the frequency feature amount calculation unit calculates at least two types of amplitude feature amounts among a plurality of types of amplitude feature amounts or a plurality of types of frequency feature amounts, and outputs the input signal of the predetermined frame. Based on the zero-crossing rate, the average value of the plurality of sample values of the input signal of the predetermined frame, or the RSM value of the plurality of sample values of the input signal of the predetermined frame, among a plurality of types of amplitude feature quantities, A feature quantity selection unit that selects an amplitude feature quantity to be calculated by the amplitude feature quantity calculation unit or a frequency feature quantity to be calculated by the frequency feature quantity calculation unit among a plurality of types of frequency feature quantities is further provided. Can do.

  The feature amount selection unit determines whether the input signal of the predetermined frame is close to a vowel or a consonant based on a zero-crossing rate of the input signal of the predetermined frame, and the amplitude according to the determination result The frequency feature quantity to be calculated by the frequency feature quantity calculation section can be selected from among the amplitude feature quantity to be calculated by the feature quantity calculation section and a plurality of types of frequency feature quantities.

  The amplitude feature amount calculation unit includes a peak value among a plurality of sample values of the predetermined frame, an average value of the plurality of sample values of the predetermined frame, or an RMS value of the plurality of sample values of the predetermined frame At least one of them as the amplitude feature amount, and the frequency feature amount calculation unit calculates a zero crossing rate of the input signal of the predetermined frame and all frequency components in the input signal of the predetermined frame. The ratio of the sound pressure of the specific frequency component to the sound pressure of the specific frequency component, the ratio of the sound pressure of the specific frequency component to the sound pressure of the frequency component different from the specific frequency component in the input signal of the predetermined frame, or At least one of a specific value or a plurality of values in a frequency spectrum obtained by Fourier transform of the input signal of the predetermined frame is the frequency characteristic. It can be made to calculate a.

  The noise determination unit includes a ratio between a weighted average value of the amplitude feature value included in the feature value set and a preset first value, and a weighted average value of the frequency feature value. 2 is calculated, a noise likelihood is calculated based on the calculated ratio, and the noise likelihood is compared with a preset threshold value, thereby obtaining the latest frame of the input signal. It can be determined whether or not the frame includes the non-stationary noise.

The noise determination unit is based on an identification model learned in advance in a feature vector space using a part or all of a weighted average value of amplitude feature amounts and a weighted average value of frequency feature amounts included in the feature amount set. Then, from the feature vector corresponding to the feature amount set, a noise likelihood representing the certainty that the frame is a non-stationary noise frame is calculated,
By comparing the noise likelihood with a preset threshold value, it can be determined whether or not the latest frame of the input signal is a frame including the non-stationary noise.

  A frequency characteristic correction unit that corrects a frequency characteristic of a signal input device that supplies the input signal may be further provided.

  A stationary noise removing unit that removes stationary noise that is different from the non-stationary noise from the input signal may be further provided.

  In one aspect of the present technology, the amplitude feature amount calculation unit calculates an amplitude feature amount in a waveform of a predetermined frame of an audio input signal, and the frequency feature amount calculation unit calculates a frequency feature amount in the waveform of the predetermined frame. The feature change amount calculating unit calculates any one of the amplitude feature amount and the frequency feature amount held in the holding unit that holds the amplitude feature amount and the frequency feature amount for a plurality of frames. And calculating a feature change amount that is a change amount of the feature amount between two temporally adjacent frames, and the section specifying unit compares the feature change amount with a preset threshold value. A frame interval in which the amplitude feature value and the frequency feature value held in the holding unit are to be weighted and averaged, and a frame segment that is temporally continuous is specified, A unit generates a set of weighted average values of the amplitude feature amount and the frequency feature amount corresponding to each of the frames in the specified section as a feature amount set, and a noise determination unit includes the feature amount set. And determining whether the latest frame of the input signal is a frame including non-stationary noise, which is sudden noise.

  One aspect of the present technology provides an amplitude feature amount calculation unit that calculates an amplitude feature amount in a waveform of a predetermined frame of an audio input signal, and a frequency feature amount that calculates a frequency feature amount in the waveform of the predetermined frame. Based on any one of the amplitude feature quantity and the frequency feature quantity held in the calculation section and the holding section that holds the amplitude feature quantity and the frequency feature quantity for a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two adjacent frames, and the holding unit by comparing the feature change amount with a preset threshold value. A section specifying unit that specifies a section of frames in which the amplitude feature quantity and the frequency feature quantity that are held are to be weighted and averaged and that is continuous in time A feature value set generation unit that generates a set of weighted average values of the amplitude feature value and the frequency feature value corresponding to each of the frames in the specified section as a feature value set; and Based on this, the program is made to function as a noise detection device including a noise determination unit that determines whether or not the latest frame of the input signal is a frame including non-stationary noise that is sudden noise.

  In one aspect of the present technology, an amplitude feature amount in a waveform of a predetermined frame of an audio input signal is calculated, a frequency feature amount in the waveform of the predetermined frame is calculated, and the amplitude feature amount and the frequency feature amount are calculated. A change in the feature amount between two temporally adjacent frames based on any one of the amplitude feature amount and the frequency feature amount held in the holding unit that holds a plurality of frames. A feature change amount that is a quantity is calculated, and the feature change amount is compared with a preset threshold value, whereby the amplitude feature amount and the frequency feature amount held in the holding unit are weighted and averaged. Sections of frames that are temporally continuous, are specified, and the amplitude feature amount corresponding to each of the frames of the specified section and the A set of weighted average values of wave number feature quantities is generated as a feature quantity set, and based on the feature quantity set, a frame including non-stationary noise in which the latest frame of the input signal is abrupt noise. It is determined whether or not there is.

  According to the present technology, various sudden noises can be detected without increasing the processing load on the device.

It is a block diagram which shows the structural example of the noise detection apparatus which concerns on one embodiment of this technique. It is a figure which shows the relationship between the frequency characteristic curve of a signal input part, and a frequency characteristic linear average. It is a block diagram which shows the detailed structural example of the frame integration part of FIG. It is a figure of the waveform which shows the waveform of an input signal, the waveform which shows the change of an amplitude feature-value, and the change of a feature-change amount. It is a flowchart explaining the example of the noise detection process by the noise detection apparatus of FIG. It is a flowchart explaining the detailed example of the integration process of FIG. It is a block diagram which shows the structural example which concerns on another embodiment of the noise detection apparatus to which this technique is applied. It is a block diagram which shows the detailed structural example of the feature-value selection part of FIG. It is a figure which shows an example of the comparison of the frequency characteristic of a cough and a vowel, and a cough and a consonant. It is a figure which shows an example of distribution of the zero crossing rate in an audio | voice signal. It is a block diagram which shows the structural example which concerns on another embodiment of the noise detection apparatus to which this technique is applied. And FIG. 16 is a block diagram illustrating a configuration example of a personal computer.

  Hereinafter, embodiments of the technology disclosed herein will be described with reference to the drawings.

  FIG. 1 is a block diagram illustrating a configuration example of a noise detection device according to an embodiment of the present technology. The noise detection apparatus 100 shown in the figure is configured to detect, for example, sudden noise (also referred to as non-stationary noise) included in surrounding sounds. Here, sudden noise is, for example, sounds such as long applause, coughing, and sneezing.

  As shown in FIG. 1, the noise detection apparatus 100 includes a frequency characteristic correction unit 101, a stationary noise reduction unit 102, an amplitude feature amount calculation unit 104, a frequency feature amount calculation unit 105, a frame integration unit 106, and a likelihood calculation unit. 107 and a noise detection unit 108.

  Further, a signal input unit 51 is connected to the noise detection device 100, and a signal processing device 52 is further connected.

  The signal input unit 51 converts a surrounding sound into a sound collecting microphone, an amplifier that amplifies a sound signal input from the microphone with an amplification factor given from the main control device, and converts an analog signal supplied from the amplifier into a digital signal. An AD converter is included.

  In recent years, a module in which an amplifier and an AD converter (including a DA converter may be integrated) has become widespread, and such a module may be provided inside the signal input unit 51. Good. The signal input unit 51 may have a function of directly reading a digital audio signal from a recording medium (for example, hard disk, CD, semiconductor memory, etc.).

For example, the frequency characteristic correction unit 101 includes a filter that interpolates the natural frequency characteristic F _id (n) of the signal input unit 51. That is, in order to prevent the digital signal supplied from the signal input unit 51 from being affected by the natural frequency characteristic of the signal input unit 51, the above-described filter removes the influence of the natural frequency characteristic of the signal input unit 51 from the input signal. Is done. Details of the processing of the frequency characteristic correction unit 101 will be described later.

  The frequency characteristic correction unit 101 supplies the signal from which the influence of the natural frequency characteristic of the signal input unit 51 is removed to the stationary noise reduction unit.

  The stationary noise reduction unit 102 calculates the stationary noise level. Here, stationary noise means noise in which frequency characteristics and amplitude characteristics included in a digital signal do not change in a long time interval. For example, the driving noise of the noise detection device 100, the signal input unit 51, or the signal processing device 52, the sound of air conditioning in the conference room, and the like are stationary noise.

  The stationary noise reduction unit 102 removes the calculated level of stationary noise component from the input signal, and then supplies it to the amplitude feature amount calculation unit 104 and the frequency feature amount calculation unit 105. For reducing stationary noise, for example, a commonly used noise reduction method may be employed, or other methods may be employed.

  The amplitude feature amount calculation unit 104 calculates one or more amplitude feature amounts from the input signal supplied from the stationary noise reduction unit 102 and supplies the calculated amount to the frame integration unit 106. Details of the amplitude feature amount will be described later.

  The frequency feature amount calculation unit 105 calculates one or more frequency feature amounts from the input signal supplied from the stationary noise reduction unit 102 and supplies the calculated frequency feature amount to the frame integration unit 106. Details of the frequency feature amount will be described later.

  The frame integration unit 106 collects a predetermined number of amplitude feature amounts and frequency feature amounts calculated for each frame supplied from the amplitude feature amount calculation unit 104 and the frequency feature amount calculation unit 105, and collects one feature amount set. Integrate as F_pack. Details of the integration method will be described later. The feature value set F_pack is supplied to the likelihood calculating unit 107.

  The likelihood calculation unit 107 calculates a ratio of each feature amount included in the feature amount set F_pack integrated by the frame integration unit 106 to a predetermined threshold value. Then, the likelihood calculating unit 107 estimates the noise likelihood for each feature amount of the feature amount set F_pack based on the calculated ratio, and calculates the weighted average value of the noise likelihood for each estimated feature amount as the noise of the input signal. Calculated as likelihood. The calculated noise likelihood is supplied to the noise detection unit 108. Details of the noise likelihood calculation method will be described later.

  The noise detection unit 108 compares the noise likelihood of the input signal supplied from the likelihood calculation unit 107 with a predetermined threshold value, and determines whether or not the input signal is nonstationary noise. The determination result by the noise detection unit 108 is output to the signal processing device 52 as the final detection result by the noise detection device 100.

  The signal processing device 52 performs signal processing using the detection result output from the noise detection unit 108. Further, the signal processing device 52 is provided with a recording unit that records an audio signal as necessary, and for example, the audio signal is recorded on a recording medium such as a hard disk, a CD, or a semiconductor memory.

  Specifically, in the signal processing device 52, for example, using the detection result output from the noise detection unit 108, the recording sensitivity adapted to only the audio portion of the input signal is calculated. For example, a recording sensitivity suitable for recording a sound excluding noise from surrounding sounds including noise is calculated.

  Further, the signal processing device 52 performs adaptive processing using the detection result output from the noise detection unit 108. For example, the signal processing device 52 uses the detection result to execute a process for reducing noise.

  Alternatively, the signal processing device 52 knows the type of noise (cough, sneeze, laughter, etc.) using the detection result, estimates the recording environment of the input signal from the type of noise, and feeds back the information. You may do it. For example, if the noise type is cough, information indicating that the health condition of the person in the recording environment is not good is fed back, and if the noise type is sneeze, the air in the place is not clean. When information is fed back and the type of noise is laughter, information indicating that the speech is interesting may be fed back.

  Next, details of the processing of the frequency characteristic correction unit 101 will be described. The frequency characteristic correction unit 101 acquires an input signal S (n) corresponding to the frame n from the signal input unit 51. Here, the input signal S (n) is defined as in Expression (1).

  In Expression (1), L is a sample value obtained as a result of sampling in AD conversion, and represents the number of sample values included in one frame, and is included in the nth frame according to Expression (1). A set of sample values to be obtained shall be obtained.

Then, the frequency characteristic correction unit 101 generates a filter H _id for correcting the natural frequency characteristic F _id (n) based on the natural frequency characteristic F _id (n) of the signal input unit 51 obtained by measurement in advance. The input signal S (n) is processed by the filter H _id so as to correct the natural frequency characteristic F _id (n) from the input signal S (n).

FIG. 2 is a diagram illustrating a relationship between a frequency characteristic curve representing a natural frequency characteristic of the signal input unit 51 and a frequency characteristic linear average that is an ideal frequency characteristic, with the vertical axis representing sound pressure and the horizontal axis representing frequency. As shown in FIG. 2, the frequency characteristic curve differs from the frequency characteristic linear average by −6 dB, +11 dB, +8 dB, and −15 dB, respectively, when the frequency is around 3 kHz, 7 kHz, 11 kHz, and 15 kHz. In this case, the frequency is 3 kHz, 7 kHz, 11 kHz, in the vicinity of 15 kHz, respectively + 6dB, -11dB, -8dB, + 15dB by only generate _{H id} to correct natural frequency characteristic from the input signal S (n) _{F id} (n ) Can be corrected to be removed.

  Note that, in the vicinity of 3 kHz, 7 kHz, 11 kHz, and 15 kHz, which are the frequencies extracted in FIG. 2, for example, the sound pressure is farthest from the frequency characteristic linear average, and is selected as a frequency that needs to be corrected. .

Alternatively, the frequency characteristic correction unit 101 generates a mapping table corresponding to the natural frequency characteristic F _id (n) of the signal input unit 51, and when calculating the amplitude feature amount and the frequency feature amount, which will be described later, The mapping table may be supplied to the amplitude feature quantity calculation unit 104 and the frequency feature quantity calculation unit 105. For example, when the frequency is around 3 kHz, 7 kHz, 11 kHz, and 15 kHz, information indicating that sound pressure is added by +6 dB, −11 dB, −8 dB, and +15 dB is used as a mapping table, and the amplitude feature amount calculation unit 104 and the frequency feature amount calculation are performed. To the unit 105.

  Note that the stationary noise reduction unit 102 may also create a mapping table in the same manner as the frequency characteristic correction unit 101 to reduce the stationary noise.

  Next, details of the amplitude feature amount will be described.

The amplitude feature quantity calculation unit 104 analyzes the amplitude characteristic of the input signal S (n) and calculates an amplitude feature quantity representing the amplitude characteristic of the frame n. Here, it is assumed that E ₁ (n), E ₂ (n), and E ₃ (n) are calculated as the amplitude feature quantities of the frame n.

E ₁ (n) is an amplitude feature amount representing the peak value of the L sample values included in the frame n, and is calculated by Expression (2).

E ₂ (n) is an amplitude feature amount that represents an average value of L sample values included in the frame n, and is calculated by Expression (3).

E ₃ (n) is an amplitude feature amount representing an RMS (Root Mean Square) value of L sample values included in the frame n, and is calculated by Expression (4).

  In addition, in Formula (3) and Formula (4), the example which calculates the linear average of a sample value was shown, For example, weighting the logarithmic average of a sample value, or the linear average and logarithmic average of a sample value A value obtained by addition may be used.

Further, before calculating E ₁ (n), E ₂ (n), and E ₃ (n), the input signal S (n) is processed by a high-pass filter to remove DC component noise contained in the input signal. You may be made to do.

Note that amplitude feature quantities other than E ₁ (n), E ₂ (n), and E ₃ (n) described above may be calculated.

  Next, details of the frequency feature amount will be described.

The frequency feature amount calculation unit 105 analyzes the frequency characteristic of the input signal S (n) and calculates a frequency feature amount representing the frequency characteristic of the frame n. Here, F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n) are calculated as the frequency feature quantities of the frame n.

F ₁ (n) is a feature amount that represents the zero-crossing rate of the input signal, and is calculated by Expression (5).

  Symbol (i) in equation (5) is expressed by equation (6).

F ₂ (n) is a feature amount that represents the ratio of the sound pressure of a specific frequency component to the sound pressure of all frequency components in the input signal, and is calculated by Expression (7).

Equation (7) _E 3 (n) in is _E 3 and (n) calculated by the equation (4).

Also, Sig _{bpf_1} (i), Sig _{bpf_2} (i),... _Shown in Expression (7) are calculated by Expression (8).

Note that F _{bpf_m} (h) in equation (8) represents a filter coefficient for extracting the m-th frequency component.

F ₃ (n) is a feature amount that represents the ratio of the sound pressure of the specific frequency component to the sound pressure of the frequency component different from the specific frequency component in the input signal, and is calculated by Expression (9). The

_{Each of} bpf _{a1_rms} (n), bpf _{a2_rms} (n), bpf _{b1_rms} (n), bpf _{b2_rms} (n), shown in equation (9) is bpf1 _rms shown as a numerator of equation (7) (N), bpf2 _rms (n),... However, when _calculating bpf _{a1_rms} (n), bpf _{a2_rms} (n), bpf _{b1_rms} (n), bpf _{b2_rms} (n),..., F _{bpf_m} (h) corresponding to each frequency component is used. And

F ₄ (n) is a feature quantity composed of a specific value or a plurality of values in the frequency spectrum obtained by Fourier transform of the input signal, and is calculated by Expression (10).

Before calculating F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n), the input signal S (n) is processed by a high-pass filter and included in the input signal. DC component noise may be removed.

Here, the amplitude feature quantity calculation unit 104 calculates E ₁ (n), E ₂ (n), and E ₃ (n), and the frequency feature quantity calculation unit 105 calculates F ₁ (n), F ₂ (n). , F ₃ (n), and F ₄ (n) are calculated. However, the amplitude feature amount calculation unit 104 calculates one or two of E ₁ (n), E ₂ (n), and E ₃ (n), and the frequency feature amount calculation unit 105 calculates F _1. Any one to three of (n), F ₂ (n), F ₃ (n), and F ₄ (n) may be calculated.

Note that frequency feature quantities other than F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n) described above may be calculated.

  Next, details of the integration method by the frame integration unit 106 will be described.

  FIG. 3 is a diagram illustrating a detailed configuration example of the frame integration unit 106. As shown in the figure, the frame integration unit 106 includes a feature holding unit 121, an integration target determination unit 122, a weight calculation unit 123, and an integration unit 124.

  The feature holding unit 121 applies the amplitude feature amount supplied from the amplitude feature amount calculation unit 104 and the frequency feature amount supplied from the frequency feature amount calculation unit 105 to the past predetermined number of frames (for example, a frame). Hold.

  The integration target determination unit 122 determines the frame to be integrated using the amplitude feature quantity or the frequency feature quantity held in the feature holding unit 121 as follows.

The integration target determination unit 122 represents a change in the feature value between frames of the feature value using any one feature value F _d of the amplitude feature value or the frequency feature value held in the feature holding unit 121. A feature change amount F _d _diff is calculated.

For example, the feature holding unit 121 may include E ₁ (n), E ₂ (n), E ₃ (n), F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n). Is maintained, E ₃ (n) is used to change the amplitude feature quantity E ₃ (i−1) of the (i−1) th frame and the amplitude feature quantity E ₃ (i) of the i th frame. The feature change amount F _d _diff representing is calculated.

The feature change amount F _d _diff is calculated by Expression (11).

The integration target determination unit 122 sequentially calculates the feature change amount between the frames using the feature amounts for all the frames held in the feature holding unit 121. Then, the calculated feature change amount is compared with a preset threshold value F _d _diff_th. In a past frame, a frame in which the feature change amount F _d _diff first exceeds the threshold value F _d _diff_th is set as an integration target start frame, and amplitude characteristics of frames (for example, b frame) from the integration target start frame to the current frame n The quantity and the frequency feature quantity are determined as integration targets. The determination result is supplied to the weight calculation unit 163.

  This will be described in more detail with reference to FIG. In FIG. 4, the horizontal axis is a frame, and in order from the top in the figure, the waveform of the input signal, the waveform indicating the change in the amplitude feature amount calculated from the input signal, and the feature change calculated based on the amplitude feature amount Each of the waveforms showing the change in quantity is shown. In the case of FIG. 4, for example, it is assumed that a coughing sound is mixed in the audio of the meeting.

  Now, the current frame is the 460th frame, and the feature holding unit 121 holds amplitude feature amounts and frequency feature amounts for 20 frames from the 441th frame to the 460th frame. And

In the example of FIG. 4, the feature change amount corresponding to the 452nd frame among the amplitude feature amounts for 20 frames first exceeds the threshold value F _d _diff_th (= 1.2). Therefore, the 452nd frame is the integration target start frame, and the nine frames up to the 460th frame are the integration target.

  In this way, a frame to be integrated is determined.

Weight calculator 163 uses the one feature F _w of the feature amounts stored in the feature storage 121, the feature amount F of other frames as a feature amount F _w of the current frame and the integration target _The weight is calculated based on the difference or ratio with _w . The weight W (i) of the i-th frame is calculated by Expression (12) or Expression (13).

Note that Equation (12) shows the expression in the case of calculating the weight based on a difference between the feature amount F _w of other frames as a feature amount F _w of the current frame and integration target, the formula (13 ) shows a formula when calculating the weight based on the ratio between the characteristic amount F _w of other frames as a feature amount F _w of the current frame and integration target.

Note that the feature value F _w used by the weight calculation unit 163 may be the same as or different from the feature value F _d used by the integration target determination unit 122.

  The weight calculated by the weight calculation unit 163 is supplied to the integration unit 124.

  The integration unit 124 uses the weight supplied from the weight calculation unit 163 to calculate the weighted average value Es (n) of the amplitude feature amount using Expression (14).

In Expression (14), n represents the current frame, and b represents the number of frames targeted for integration. In addition, as described above, when a plurality of amplitude feature quantities (for example, E ₁ (n), E ₂ (n), and E ₃ (n)) are held in the feature holding unit 121, Expression (14) E (n) in E _{2 is} E ₁ (n), E ₂ (n), and E ₃ (n), respectively, and the weighted average value Es ₁ (n) to weighted average value Es ₃ (n ) Are respectively calculated.

  Further, the integration unit 124 calculates the weighted average value Fs (n) of the frequency feature amount by using the weight supplied from the weight calculation unit 163 according to Expression (15).

In equation (15), n represents the current frame, and b represents the number of frames targeted for integration. Further, as described above, a plurality of frequency feature quantities (for example, F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n)) are held in the feature holding unit 121. In this case, F (n) in Expression (14) is set to F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n), respectively, and the weighted average value Fs ₁ of the frequency feature amount (N) to Fs ₄ (n) are respectively calculated.

  Then, the integrating unit 124 supplies a set of the weighted average value Es (n) of the amplitude feature quantity and the weighted average value Fs (n) of the frequency feature quantity to the likelihood calculating unit 107 as a feature quantity set F_pack.

  The frame integration unit 106 does not include the weight calculation unit 163, and the integration unit 124 sets a simple average set of the amplitude feature amount and the frequency feature amount of the integration target frame determined by the integration target determination unit 122. The feature amount set F_pack may be generated by integration.

  Further, the integration unit determining unit 162 is not included in the frame integration unit 106, the weight calculation unit 123 calculates the weights of all frames held by the feature holding unit 121, and the integration unit 124 calculates the amplitudes of all frames. You may make it produce | generate the feature-value set F_pack which integrated the set of the weighted average of a feature-value and a frequency feature-value.

  Furthermore, the integration unit determination unit 162 and the weight calculation unit 163 are not included in the frame integration unit 106, and the integration unit 124 simply calculates the amplitude feature amounts and frequency feature amounts of all frames held by the feature holding unit 121. A set of average values may be generated as a feature value set F_pack.

  The likelihood calculation unit 107 calculates a ratio of each feature amount included in the feature amount set F_pack integrated by the frame integration unit 106 to a predetermined threshold value.

  For example, a threshold value E_th corresponding to the amplitude feature value and a threshold value F_th corresponding to the frequency feature value are determined in advance.

The likelihood calculating unit 107 calculates the ratio R _E (n) of the threshold value E_th with respect to the weighted average value of the amplitude feature amount included in the feature amount set F_pack, using Expression (16).

In addition, the likelihood calculating unit 107 calculates the ratio R _F (n) of the threshold value F_th with respect to the weighted average value of the frequency feature amounts included in the feature amount set F_pack, using Expression (17).

The likelihood calculating unit 107 calculates a weighted addition value by multiplying each of the ratio R _E (n) and the ratio R _F (n) by a predetermined weight A _E and weight A _F. This weighted addition value is calculated by Expression (18), and is supplied to the noise detection unit 108 as the noise likelihood R (n) corresponding to the nth frame of the input signal.

  The noise detection unit 108 compares the noise likelihood of the input signal supplied from the likelihood calculation unit 107 with a predetermined threshold value, and determines whether or not the nth frame of the input signal is a frame of nonstationary noise. Determine. For example, when the noise likelihood threshold R_th for determining non-stationary noise is determined in advance and the noise likelihood R (n) is larger than the noise likelihood threshold R_th, the nth frame of the input signal is not non-stationary. It is determined that the frame is stationary noise. On the other hand, when the noise likelihood R (n) is equal to or less than the noise likelihood threshold R_th, it is determined that the nth frame of the input signal is not a frame of nonstationary noise.

  In this way, non-stationary noise is detected. In the present technology, as described above, since it is determined whether or not it is non-stationary noise using at least one amplitude feature quantity and at least one frequency feature quantity, non-stationary noise is determined. Can be detected with higher accuracy.

  In addition, since the frame to be integrated is specified in the frame integration unit 106, it is possible to reduce the calculation load of the feature amount included in the feature amount set F_pack. Thereby, for example, the noise detection apparatus 100 can be mounted on a small power-saving device.

  Furthermore, by setting the noise likelihood threshold as a dedicated noise likelihood threshold for detecting cough, it is possible to determine only cough as non-stationary noise, and a dedicated noise likelihood for detecting applause. By setting the threshold value, only applause can be determined as non-stationary noise. Thus, in the present technology, it is possible to specify the type of non-stationary noise by appropriately setting the noise likelihood threshold.

  In the example described above, the likelihood calculating unit 107 performs threshold comparison based on the threshold E_th corresponding to the preset amplitude feature quantity and the threshold F_th corresponding to the frequency feature quantity, and Expressions (16) to (18) are performed. The noise likelihood was calculated by performing the above calculation.

  However, for example, the likelihood calculating unit 107 may calculate the noise likelihood from the feature amount set F_pack using the identification model M learned in advance. In this case, for example, a Gaussian mixture model (GMM), a hidden Markov model (HMM), a support vector machine (SVM), or the like can be adopted as the identification model M.

  That is, a feature vector space is generated using part or all of the weighted average value of amplitude feature values and the weighted average value of frequency feature values included in the feature value set F_pack. Based on the identification model learned in advance in the feature vector space, the likelihood calculation unit 107 determines the certainty that the frame is a frame of nonstationary noise from the feature vector corresponding to the feature amount set F_pack. The noise likelihood to represent is calculated.

  Note that the likelihood calculation method using these identification models is the same as that generally employed conventionally.

  Next, an example of noise detection processing by the noise detection apparatus 100 will be described with reference to the flowchart of FIG.

  In step S <b> 21, the frequency characteristic correction unit 101 acquires the input signal S (n) output from the signal input unit 51.

In step S < _b > 22, the frequency characteristic correction unit 101 corrects the natural frequency characteristic F _id (n) of the signal input unit 51. At this time, for example, the natural frequency characteristic as described above with reference to FIG. 2 is corrected, and the influence of the natural frequency characteristic of the signal input unit 51 is removed from the input signal.

  In step S23, the stationary noise reduction unit 102 removes stationary noise. Thereby, for example, the driving sound of the noise detection device 100, the signal input unit 51, or the signal processing device 52, the sound of air conditioning in the conference room, and the like are removed.

In step S <b> 24, the amplitude feature amount calculation unit 104 calculates the amplitude feature amount from the input signal supplied from the stationary noise reduction unit 102. At this time, at least one of E ₁ (n), E ₂ (n), and E ₃ (n) described above is calculated as the amplitude feature quantity of frame n.

In step S <b> 25, the frequency feature amount calculation unit 105 calculates a frequency feature amount from the input signal supplied from the stationary noise reduction unit 102. At this time, at least one of the above-described F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n) is calculated as the frequency feature amount of the frame n.

  In step S26, the frame integration unit 106 executes integration processing to be described later with reference to FIG. As a result, the amplitude feature quantity calculated in the process of step S24 and the frequency feature quantity calculated in the process of step S25 are integrated for a predetermined number of frames, and the weighted average value Es (n) and the frequency of the amplitude feature quantity are integrated. A weighted average value Fs (n) of feature amounts is calculated. A set of the weighted average value Es (n) of the amplitude feature quantity and the weighted average value Fs (n) of the frequency feature quantity is output as a feature quantity set F_pack.

In step S27, the likelihood calculation unit 107 calculates the noise likelihood of the input signal. At this time, as described above, the ratio of the threshold value E_th corresponding to the amplitude feature value and the threshold value F_th corresponding to the frequency feature value is calculated for each feature value included in the feature value set F_pack. Then, each of the ratio R _E (n) and the ratio R _F (n) is multiplied by a predetermined weight A _E and weight A _F to calculate a weighted addition value, and the nth frame of the input signal is calculated. The corresponding noise likelihood is R (n).

  In step S28, the noise detection unit 108 determines whether or not the noise likelihood R (n) is larger than the noise likelihood threshold R_th.

  In Step S28, when it is determined that the noise likelihood R (n) is larger than the noise likelihood threshold R_th, the process proceeds to Step S29.

  In step S29, the noise detection unit 108 determines that the nth frame of the input signal is a frame of nonstationary noise.

  On the other hand, when it is determined in step S28 that the noise likelihood R (n) is not larger than the noise likelihood threshold R_th, the process proceeds to step S30.

  In step S30, the noise detection unit 108 determines that the nth frame of the input signal is not a non-stationary noise frame.

  In this way, the noise detection process is executed.

  Next, a detailed example of the integration process in step S26 in FIG. 5 will be described with reference to the flowchart in FIG.

  In step S <b> 51, the integration target determination unit 122 acquires the amplitude feature quantity and the frequency feature quantity held in the feature holding unit 121.

In step S52, the integration target determination unit 122, using any one of the feature amount F _d of the amplitude characteristic quantity or frequency feature amount acquired in step S51, indicating a change in the characteristic amount between the feature quantity of the frame A feature change amount F _d _diff is calculated. The feature change amount F _d _diff is calculated for all the frames corresponding to the amplitude feature amount and the frequency feature amount held in the feature holding unit 121.

For example, the feature holding unit 121 may include E ₁ (n), E ₂ (n), E ₃ (n), F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n). Is maintained, E ₃ (n) is used to change the amplitude feature quantity E ₃ (i−1) of the (i−1) th frame and the amplitude feature quantity E ₃ (i) of the i th frame. A feature change amount F _d _diff (i) representing is calculated.

  In step S53, the integration target determination unit 122 sets a number n representing the current frame to the variable i.

In step S54, the integration target determining unit 122 compares the feature change amount F _d _diff (i) with a preset threshold value F _d _diff_th, and whether the feature change amount F _d _diff (i) exceeds the threshold value F _d _diff_th. Determine whether or not.

If it is determined in step S54 that the feature change amount F _d _diff (i) does not exceed the threshold value F _d _diff_th, the process proceeds to step S55.

  In step S55, the variable i is decremented, and the process returns to step S54.

On the other hand, when it is determined in step S54 that the feature change amount F _d _diff (i) exceeds the threshold value F _d _diff_th, the process proceeds to step S56.

  In step S56, the integration target determination unit 122 determines the i-th frame (i frame) to the n-th frame (n frame) as integration targets. In this case, the i frame is the integration target start frame.

In step S57, the weight calculation section 163, using a single feature value F _w of the feature amounts stored in the feature storage 121, the current frame feature value F _w integrated subject to other frames calculating a weighting based on the difference or ratio between the characteristic amount F _w. Note that the feature value F _w used by the weight calculation unit 163 may be the same as or different from the feature value F _d used by the integration target determination unit 122.

  The weight calculated by the weight calculation unit 163 is supplied to the integration unit 124.

  The integration unit 124 uses the weight supplied from the weight calculation unit 163 to calculate the weighted average value Es (n) of the amplitude feature amount using Expression (14).

  In step S58, the integration unit 124 calculates the weighted average value Es (n) of the amplitude feature quantity and the weighted average value Fs (n) of the frequency feature quantity using the weight calculated in the process of step S57.

  In step S59, the integration unit 124 generates a set of the weighted average value Es (n) of the amplitude feature quantity and the weighted average value Fs (n) of the frequency feature quantity as the feature quantity set F_pack.

  In this way, the integration process is executed.

  FIG. 7 is a block diagram illustrating a configuration example according to another embodiment of the noise detection apparatus 100 to which the present technology is applied. Unlike the case of FIG. 1, the noise detection device 100 in the configuration of FIG. 7 includes a feature amount selection unit 103. The other configuration of the noise detection apparatus 100 of FIG. 7 is the same as that of FIG.

  The feature quantity selection unit 103 calculates the amplitude feature quantity to be calculated by the amplitude feature quantity calculation unit 104 and the frequency feature quantity calculation unit 105 based on the input signal output through the processing of the stationary noise reduction unit 102. Specify the frequency feature to be used. Thereby, the calculation load of the amplitude feature quantity calculation unit 104 and the frequency feature quantity calculation unit 105 can be reduced.

  FIG. 8 is a block diagram illustrating a detailed configuration example of the feature amount selection unit 103. As shown in the figure, the feature amount selection unit 103 includes a feature calculation unit 131, a feature determination unit 132, and a selection information output unit 133.

The feature calculation unit 131 calculates the feature amount of the input signal and supplies it to the feature determination unit 132. The feature amount calculated by the feature calculation unit 131 is, for example, E ₁ (n), E ₂ (n), and E ₃ (n), which are the above-described amplitude feature amounts, or the above-described frequency feature amount. One of F ₁ (n), F ₂ (n), F ₃ (n), and F ₄ (n).

  The feature determination unit 132 compares the feature amount supplied from the feature calculation unit 131 with a threshold, determines the feature type of the input signal of the frame from the result, and supplies the feature type to the selection information output unit 133. .

  The selection information output unit 133 uses the feature type supplied from the feature determination unit 132 to select feature selection information corresponding to each feature type, and uses the feature selection information as the amplitude feature quantity calculation unit 104 and the frequency feature. It outputs to the quantity calculation part 105. Here, the feature selection information is information specifying the amplitude feature amount to be calculated by the amplitude feature amount calculation unit 104 and the frequency feature amount to be calculated by the frequency feature amount calculation unit 105.

  FIG. 9 is a diagram for explaining the frequency characteristics of cough, which is one of non-stationary noises, and is a diagram showing an example of comparison of frequency characteristics of cough and vowels and cough and consonants. In the figure, the horizontal axis represents frequency, the vertical axis represents sound pressure level, and the frequency characteristics related to cough speech and the frequency characteristics related to normal speech are indicated by broken lines. The upper side of the figure shows the frequency characteristics of the vowel voice, the cough voice, and the cough voice, and the lower side of the figure shows the frequency characteristics of the consonant voice and the cough voice. .

  As shown in the upper part of the figure, when comparing cough voice and vowel voice, the sound pressure in the section below 1.4kHz, the section from 4kHz to 6.8kHz, and the section above 11.7kHz. The levels are very different. In other words, the frequency feature quantity of these sections, for example, a frequency band component of 1.4 kHz or less, a frequency band component of 4 kHz to 6.8 kHz, and a frequency band component of 11.7 kHz or more are used to extract the input signal. If a set of parameters indicating the ratio of the frequency components in the above-described section to all frequency components is calculated, cough speech and vowel speech can be easily distinguished.

  In addition, as shown in the lower part of the figure, when comparing cough voice and consonant voice, the section below 1.8 kHz, the section from 6.5 kHz to 8.8 kHz, and the section above 17.7 kHz The sound pressure level varies greatly between sections. That is, the cough voice and the consonant voice can be easily distinguished by using a filter that extracts the frequency band component of each section as in the case of the comparison between the cough voice and the vowel voice.

  However, in order to compare cough and vowel and cough and consonant, it is necessary to extract different frequency components, and in order to detect cough with high accuracy, feature values related to a total of six frequency components are calculated. There is a need to. That is, if it is not known in advance whether the input signal is a sound close to a vowel or a sound close to a consonant, the feature amount must be calculated assuming both cases.

  For example, if it is possible to recognize in advance whether the input signal is a sound close to a vowel or a sound close to a consonant, it is only necessary to calculate the feature amounts relating to a total of three frequency components. It is possible to reduce the load related to the amount calculation.

  FIG. 10 is a diagram illustrating an example of the distribution of the zero-crossing rate of the audio signal obtained as a result of an experiment using a plurality of audio signals as samples. In the figure, the horizontal axis represents the zero-crossing rate, and the vertical axis represents the number of samples per frame of the audio signal having the zero-crossing rate.

  As shown in FIG. 10, two Gaussian characteristics can be seen in the sample distribution with a zero crossing rate of 0.05 as a boundary. It has been found that most of the samples having a zero crossing rate of 0.05 or less are vowels. On the other hand, it is known that most of the samples having a zero crossing rate of 0.05 or more are consonants.

  That is, by setting the zero crossing rate 0.05 as the threshold value F_th and comparing the zero crossing rate of the input signal with the threshold value F_th, it is recognized whether the input signal is a sound close to a vowel or a sound close to a consonant. It becomes possible.

  For example, the feature calculation unit 131 of the feature amount selection unit 103 calculates the zero-crossing rate of the input signal, and the feature determination unit 132 compares the zero-crossing rate of the input signal with the threshold value F_th. It is determined whether the feature type of the input signal is a vowel or a consonant. Thus, the amplitude feature quantity to be calculated by the amplitude feature quantity calculation unit 104 and the frequency feature quantity to be calculated by the frequency feature quantity calculation unit 105 are set as a vowel feature quantity or a consonant feature quantity.

  Thus, by providing the feature quantity selection unit 103, the calculation load of the amplitude feature quantity calculation unit 104 and the frequency feature quantity calculation unit 105 can be reduced.

  Here, an example has been described in which the feature amount selection unit 103 determines whether the feature type of the input signal of the frame is a vowel or a consonant. For example, the feature type of the input signal of the frame is a sound. It may be determined whether the pressure is high (high sound pressure) or low (low sound pressure). For example, in the case of a low sound pressure (when the volume is low), it is difficult to obtain a good S / N characteristic, so that a feature amount that is not easily affected by stationary noise may be selected.

In this case, instead of the zero crossing rate, the amplitude feature amount (E ₂ (n)) representing the average value of the L sample values included in the frame n, or the RMS of the L sample values included in the frame n The feature type of the input signal of the frame may be determined by comparing the amplitude feature amount (E ₃ (n)) representing the value with a threshold value.

  FIG. 11 is a block diagram illustrating a configuration example according to still another embodiment of the noise detection apparatus 100 to which the present technology is applied. Unlike the case of FIG. 1, the noise detection apparatus 100 in the configuration of FIG. 11 does not include the frequency characteristic correction unit 101, the stationary noise reduction unit 102, the frame integration unit 106, and the likelihood calculation unit 107. The other configuration of the noise detection apparatus 100 of FIG. 11 is the same as that of FIG.

  In the case of the configuration of FIG. 11, the noise detection apparatus 100 directly calculates the amplitude feature quantity and the frequency feature quantity from the input signal supplied from the signal input unit 51 and directly uses the amplitude feature quantity and the frequency feature quantity. Thus, it is determined whether or not the frame is a non-stationary noise frame. In this case, for example, the noise detection unit 108 performs threshold determination on each of the amplitude feature amount and the frequency feature amount, and determines whether or not the frame is a frame of nonstationary noise corresponding to the determination result. become.

  Alternatively, any one to three of the frequency characteristic correction unit 101, the stationary noise reduction unit 102, the frame integration unit 106, and the likelihood calculation unit 107 are added to the noise detection device 100 illustrated in FIG. It is also possible to adopt an additional configuration.

  The series of processes described above can be executed by hardware, or can be executed by software. When the above-described series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, a general-purpose personal computer 700 as shown in FIG. 12 is installed from a network or a recording medium.

  In FIG. 12, a CPU (Central Processing Unit) 701 executes various processes according to a program stored in a ROM (Read Only Memory) 702 or a program loaded from a storage unit 708 to a RAM (Random Access Memory) 703. To do. The RAM 703 also appropriately stores data necessary for the CPU 701 to execute various processes.

  The CPU 701, ROM 702, and RAM 703 are connected to each other via a bus 704. An input / output interface 705 is also connected to the bus 704.

  The input / output interface 705 includes an input unit 706 including a keyboard and a mouse, a display including an LCD (Liquid Crystal display), an output unit 707 including a speaker, a storage unit 708 including a hard disk, a modem, a LAN, and the like. A communication unit 709 including a network interface card such as a card is connected. The communication unit 709 performs communication processing via a network including the Internet.

  A drive 710 is also connected to the input / output interface 705 as necessary, and a removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory is appropriately mounted, and a computer program read from them is loaded. It is installed in the storage unit 708 as necessary.

  When the above-described series of processing is executed by software, a program constituting the software is installed from a network such as the Internet or a recording medium such as a removable medium 711.

  The recording medium shown in FIG. 12 is a magnetic disk (including a floppy disk (registered trademark)) on which a program is recorded, which is distributed to distribute the program to the user separately from the apparatus main body. Removable media consisting of optical disks (including CD-ROM (compact disk-read only memory), DVD (digital versatile disk)), magneto-optical disks (including MD (mini-disk) (registered trademark)), or semiconductor memory It includes not only those configured by 711 but also those configured by a ROM 702 in which a program is recorded, a hard disk included in the storage unit 708, and the like distributed to the user in a state of being incorporated in the apparatus main body in advance.

  Note that the series of processes described above in this specification includes processes that are performed in parallel or individually even if they are not necessarily processed in time series, as well as processes that are performed in time series in the order described. Is also included.

  The embodiments of the present technology are not limited to the above-described embodiments, and various modifications can be made without departing from the gist of the present technology.

  In addition, this technique can also take the following structures.

(1)
An amplitude feature amount calculation unit for calculating an amplitude feature amount in a waveform of a predetermined frame of an audio input signal;
A frequency feature amount calculation unit for calculating a frequency feature amount in the waveform of the predetermined frame;
2 adjacent in terms of time based on any one of the amplitude feature quantity and the frequency feature quantity held in the holding section that holds the amplitude feature quantity and the frequency feature quantity for a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two frames;
By comparing the feature change amount with a preset threshold, the amplitude feature amount and the frequency feature amount held in the holding unit are sections of frames to be weighted and averaged, and are temporally continuous. A section identifying unit that identifies a section of the frame;
A feature quantity set generation unit that generates a set of weighted average values of the amplitude feature quantity and the frequency feature quantity corresponding to each of the frames of the specified section as a feature quantity set;
A noise detection apparatus comprising: a noise determination unit that determines whether the latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set.
(2)
The amplitude feature amount calculation unit or the frequency feature amount calculation unit calculates at least two types of amplitude feature amounts among a plurality of types of amplitude feature amounts or a plurality of types of frequency feature amounts,
Based on a zero-crossing rate of the input signal of the predetermined frame, an average value of a plurality of sample values of the input signal of the predetermined frame, or a plurality of RSM values of a plurality of sample values of the input signal of the predetermined frame Feature quantity selection for selecting an amplitude feature quantity to be calculated by the amplitude feature quantity calculation unit from among the types of amplitude feature quantities, or a frequency feature quantity to be calculated by the frequency feature quantity calculation unit from among a plurality of types of frequency feature quantities The noise detection device according to (1), further including a unit.
(3)
The feature amount selection unit includes:
Based on the zero-crossing rate of the input signal of the predetermined frame, it is determined whether the input signal of the predetermined frame is close to a vowel or a consonant, and the amplitude feature amount calculation unit is made to calculate according to the determination result The noise detection apparatus according to (2), wherein a frequency feature amount to be calculated by the frequency feature amount calculation unit is selected from among an amplitude feature amount and a plurality of types of frequency feature amounts.
(4)
The amplitude feature amount calculation unit includes:
At least one of a peak value among a plurality of sample values of the predetermined frame, an average value of the plurality of sample values of the predetermined frame, or an RMS value of the plurality of sample values of the predetermined frame Calculated as the amplitude feature amount,
The frequency feature amount calculation unit includes:
The zero-crossing rate of the input signal of the predetermined frame, the ratio of the sound pressure of a specific frequency component to the sound pressure of all frequency components in the input signal of the predetermined frame, and the input signal of the predetermined frame The ratio of the sound pressure of the specific frequency component to the sound pressure of the frequency component different from the specific frequency component, or a specific one of the frequency spectrum obtained by Fourier transforming the input signal of the predetermined frame The noise detection device according to any one of (1) to (3), wherein at least one of a value or a plurality of values is calculated as the frequency feature amount.
(5)
The noise determination unit
The ratio between the weighted average value of the amplitude feature quantity included in the feature quantity set and a preset first value, and the ratio between the weighted average value of the frequency feature quantity and a preset second value To calculate
Based on the calculated ratio, a noise likelihood is calculated,
It is determined whether the latest frame of the input signal is a frame including the non-stationary noise by comparing the noise likelihood with a preset threshold value (1) to (4) The noise detection apparatus described in 1.
(6)
The noise determination unit
In the feature vector space using part or all of the weighted average value of the amplitude feature quantity and the weighted average value of the frequency feature quantity included in the feature quantity set, the feature quantity set is based on a previously learned identification model. From the feature vector corresponding to, a noise likelihood representing the probability that the frame is a non-stationary noise frame is calculated,
It is determined whether the latest frame of the input signal is a frame including the non-stationary noise by comparing the noise likelihood with a preset threshold value. The noise detection apparatus described in 1.
(7)
The noise detection device according to any one of (1) to (6), further including a frequency characteristic correction unit that corrects a frequency characteristic of a signal input device that supplies the input signal.
(8)
The noise detection apparatus according to any one of (1) to (7), further including a stationary noise removing unit that removes stationary noise that is different from the non-stationary noise from the input signal.
(9)
The amplitude feature amount calculation unit calculates the amplitude feature amount in the waveform of a predetermined frame of the voice input signal,
The frequency feature amount calculation unit calculates a frequency feature amount in the waveform of the predetermined frame,
The feature change amount calculation unit is based on any one of the amplitude feature amount and the frequency feature amount held in the holding unit that holds the amplitude feature amount and the frequency feature amount for a plurality of frames. Calculating a feature change amount that is a change amount of the feature amount between two temporally adjacent frames;
A section specifying unit is a section of a frame in which the amplitude feature amount and the frequency feature amount held in the holding unit are weighted and averaged by comparing the feature change amount with a preset threshold value, Identify the interval between successive frames,
A feature amount set generation unit generates a set of weighted average values of the amplitude feature amount and the frequency feature amount corresponding to each of the frames of the specified section as a feature amount set;
A noise detection method comprising: a step of determining whether or not a latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set.
(10)
Computer
An amplitude feature amount calculation unit for calculating an amplitude feature amount in a waveform of a predetermined frame of an audio input signal;
A frequency feature amount calculation unit for calculating a frequency feature amount in the waveform of the predetermined frame;
2 adjacent in terms of time based on any one of the amplitude feature quantity and the frequency feature quantity held in the holding section that holds the amplitude feature quantity and the frequency feature quantity for a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two frames;
By comparing the feature change amount with a preset threshold, the amplitude feature amount and the frequency feature amount held in the holding unit are sections of frames to be weighted and averaged, and are temporally continuous. A section identifying unit that identifies a section of the frame;
A feature quantity set generation unit that generates a set of weighted average values of the amplitude feature quantity and the frequency feature quantity corresponding to each of the frames of the specified section as a feature quantity set;
A program that functions as a noise detection device including a noise determination unit that determines whether the latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set .

  51 signal input unit, 52 signal processing device, 100 noise detection device, 101 frequency characteristic correction unit, 102 stationary noise reduction unit, 103 feature amount selection unit, 104 amplitude feature amount calculation unit, 105 frequency feature amount calculation unit, 106 frame Integration unit, 107 likelihood calculation unit, 108 noise detection unit, 121 feature holding unit, 122 integration target determination unit, 123 weight calculation unit, 124 integration unit, 131 feature calculation unit, 132 feature determination unit, 133 selection information output unit, 711 Removable media

Claims (10)

An amplitude feature amount calculation unit for calculating an amplitude feature amount in a waveform of a predetermined frame of an audio input signal;
A frequency feature amount calculation unit for calculating a frequency feature amount in the waveform of the predetermined frame;
2 adjacent in terms of time based on any one of the amplitude feature quantity and the frequency feature quantity held in the holding section that holds the amplitude feature quantity and the frequency feature quantity for a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two frames;
By comparing the feature change amount with a preset threshold, the amplitude feature amount and the frequency feature amount held in the holding unit are sections of frames to be weighted and averaged, and are temporally continuous. A section identifying unit that identifies a section of the frame;
A feature quantity set generation unit that generates a set of weighted average values of the amplitude feature quantity and the frequency feature quantity corresponding to each of the frames of the specified section as a feature quantity set;
A noise detection apparatus comprising: a noise determination unit that determines whether the latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set. The amplitude feature amount calculation unit or the frequency feature amount calculation unit calculates at least two types of amplitude feature amounts among a plurality of types of amplitude feature amounts or a plurality of types of frequency feature amounts,
Based on a zero-crossing rate of the input signal of the predetermined frame, an average value of a plurality of sample values of the input signal of the predetermined frame, or a plurality of RSM values of a plurality of sample values of the input signal of the predetermined frame Feature quantity selection for selecting an amplitude feature quantity to be calculated by the amplitude feature quantity calculation unit from among the types of amplitude feature quantities, or a frequency feature quantity to be calculated by the frequency feature quantity calculation unit from among a plurality of types of frequency feature quantities The noise detection apparatus according to claim 1, further comprising a unit. The feature amount selection unit includes:
Based on the zero-crossing rate of the input signal of the predetermined frame, it is determined whether the input signal of the predetermined frame is close to a vowel or a consonant, and the amplitude feature amount calculation unit is made to calculate according to the determination result The noise detection device according to claim 2, wherein a frequency feature amount to be calculated by the frequency feature amount calculation unit is selected from an amplitude feature amount and a plurality of types of frequency feature amounts. The amplitude feature amount calculation unit includes:
At least one of a peak value among a plurality of sample values of the predetermined frame, an average value of the plurality of sample values of the predetermined frame, or an RMS value of the plurality of sample values of the predetermined frame Calculated as the amplitude feature amount,
The frequency feature amount calculation unit includes:
The zero-crossing rate of the input signal of the predetermined frame, the ratio of the sound pressure of a specific frequency component to the sound pressure of all frequency components in the input signal of the predetermined frame, and the input signal of the predetermined frame The ratio of the sound pressure of the specific frequency component to the sound pressure of the frequency component different from the specific frequency component, or a specific one of the frequency spectrum obtained by Fourier transforming the input signal of the predetermined frame The noise detection device according to claim 1, wherein at least one of a value or a plurality of values is calculated as the frequency feature amount. The noise determination unit
The ratio between the weighted average value of the amplitude feature quantity included in the feature quantity set and a preset first value, and the ratio between the weighted average value of the frequency feature quantity and a preset second value To calculate
Based on the calculated ratio, a noise likelihood is calculated,
The noise detection device according to claim 1, wherein the noise likelihood is determined by comparing the noise likelihood with a preset threshold value to determine whether the latest frame of the input signal is a frame including the non-stationary noise. The noise determination unit
In the feature vector space using part or all of the weighted average value of the amplitude feature quantity and the weighted average value of the frequency feature quantity included in the feature quantity set, the feature quantity set is based on a previously learned identification model. From the feature vector corresponding to, a noise likelihood representing the probability that the frame is a non-stationary noise frame is calculated,
The noise detection device according to claim 1, wherein the noise likelihood is determined by comparing the noise likelihood with a preset threshold value to determine whether the latest frame of the input signal is a frame including the non-stationary noise. The noise detection apparatus according to claim 1, further comprising a frequency characteristic correction unit that corrects a frequency characteristic of a signal input device that supplies the input signal. The noise detection apparatus according to claim 1, further comprising a stationary noise removing unit that removes stationary noise that is different from the non-stationary noise from the input signal. The amplitude feature amount calculation unit calculates the amplitude feature amount in the waveform of a predetermined frame of the voice input signal,
The frequency feature amount calculation unit calculates a frequency feature amount in the waveform of the predetermined frame,
The feature change amount calculation unit is based on any one of the amplitude feature amount and the frequency feature amount held in the holding unit that holds the amplitude feature amount and the frequency feature amount for a plurality of frames. Calculating a feature change amount that is a change amount of the feature amount between two temporally adjacent frames;
A section specifying unit is a section of a frame in which the amplitude feature amount and the frequency feature amount held in the holding unit are weighted and averaged by comparing the feature change amount with a preset threshold value, Identify the interval between successive frames,
A feature amount set generation unit generates a set of weighted average values of the amplitude feature amount and the frequency feature amount corresponding to each of the frames of the specified section as a feature amount set;
A noise detection method comprising: a step of determining whether or not a latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set. Computer
An amplitude feature amount calculation unit for calculating an amplitude feature amount in a waveform of a predetermined frame of an audio input signal;
A frequency feature amount calculation unit for calculating a frequency feature amount in the waveform of the predetermined frame;
2 adjacent in terms of time based on any one of the amplitude feature quantity and the frequency feature quantity held in the holding section that holds the amplitude feature quantity and the frequency feature quantity for a plurality of frames. A feature change amount calculation unit that calculates a feature change amount that is a change amount of the feature amount between two frames;
By comparing the feature change amount with a preset threshold, the amplitude feature amount and the frequency feature amount held in the holding unit are sections of frames to be weighted and averaged, and are temporally continuous. A section identifying unit that identifies a section of the frame;
A feature quantity set generation unit that generates a set of weighted average values of the amplitude feature quantity and the frequency feature quantity corresponding to each of the frames of the specified section as a feature quantity set;
A program that functions as a noise detection device including a noise determination unit that determines whether the latest frame of the input signal is a frame including non-stationary noise that is sudden noise based on the feature amount set .

Images